| Different from the traditional single-label classification problem,in a multi-label classification problem,one sample is usually defined by multiple labels.Due to the rapid development of data collection technology and the diversification of data storage technologies,multi-label data is widespread,and the multi-label classification problem is more in line with people’s understanding of actual problems,so the multi-label classification problem has been widely studied and applied.Since the multi-label classification problem,many classic algorithms have emerged.One is the problem transformation method that converts a multi-label data set into a single-label data set and then uses a single-label classification algorithm.The other one is the algorithm adaptive method that applied to the multi-label data set based on the single-label classification algorithm.Due to the high feature dimension of multi-label data sets,most of them have redundant and irrelevant features,which leads to the increase of training complexity of multi-label classification algorithm and the decrease of prediction accuracy.Therefore,this paper considers feature selection to identify and remove redundant features and select a good feature set.Different from the feature selection of the single label classification problem,there is correlation between label spaces,which makes many filtering feature selection algorithms based on data conversion algorithms that do not consider label correlation have limited improvement in classification effect.This paper focuses on the correlation between labels,and uses the Principal Component Analysis to map the original label space to the primary label space.And then,there is no correlation between the primary labels any more,and setting weights according to their respective variance contribution rate is conducive to the reasonable assessment of the correlation between features and label space.Based on this idea,this paper modified the candidate subset evaluation function for the ML-CFS algorithm and proposed the ML-PLCFS algorithm.In order to compare the two algorithms more precisely,the improved Genetic Algorithm is used to control the size of feature subsets and analyze and compare the simulated multi-label data sets and the real multi-label application scenarios in four different fields.In this paper,5 commonly used evaluation metrics are selected and the subset score and effective feature proportion are customized.Experimental results show that the ML-PLCFS algorithm proposed in this paper is significantly superior to the ML-CFS algorithm. |
